PARIS — PyTorch’s co-founder told Open Source AI Developers last week that the project explored, but ultimately abandoned, tighter integration of its AI framework with languages such as Rust and JavaScript.
“PyTorch is like a typical hacker who will take any shortcut to give good performance to the users. We are very pragmatic,” Soumith Chintala explained the ethos behind the framework during a fireside chat at the Linux Foundation’s AI Developers Conference here.
When asked about the future roadmap for PyTorch, he responded, “Tools need to evolve, but they don’t need to evolve very often.”
When it comes to the AI field, “I think every six or seven years, new tools come out,” he said, so “I don’t see the need for PyTorch to even think about 2.0 anytime soon. But we’re continuing to look at it, and things are continuing to change.”
After ChatGPT launched, Transformers became by far the primary way to run AI, and it’s probably “something we need to think about carefully, but overall we’re still pretty happy with it,” he said.
Integration with other languages?
One thing that’s obviously not on the agenda is tight integration with languages other than Python.
In response to a question from the audience asking “Will PyTorch be offered for Rust or JavaScript?”, his answer was “PyTorch has ‘Py’ in it, so it won’t be officially offered.”
Chintala added that the team had looked into it, but the problem always came with the compiler, he continued: “You have to parse the front-end code, and writing a front-end that uses user code in different languages was a lot of work.
“From a priorities and research standpoint, we decided no, but we don’t actually have any animosity towards Rust or JavaScript.”
Docker and WebGPU
Elsewhere at the event, Docker previewed WebGPU support in Docker, announcing that it will eliminate the need to build Docker images specifically for each GPU vendor’s proprietary drivers.
Justin Cormack, Docker’s UK-based CTO, says that when the company was founded, developer hardware environments were “fairly homogenous. You could pretty much ship a laptop straight into production.”
But the explosion of AI and reliance on GPUs has changed that, Cormack continued. “We’re seeing a wide range of different accelerators being used on development machines, on edge machines, and in production,” he added. “There’s a huge variety.”
That made it “a really exciting time for hardware,” he says, but inevitably developers were “asking about using GPUs.” The ideal was a portable abstraction with Docker that could be used anywhere, he says. And the answer came in the form of WebGPU, a World Wide Web Consortium (W3C)-backed API for running operations on GPUs.
“The impression I got was that WebGPU already exists, it’s already in all consumer GPUs and most browsers on mobile, and it has an ecosystem,” he said. But he added that WebGPU doesn’t just work in browsers because it has a portable standard definition and a more performant modern stack compared to WebGL.
Cormack said the company is shipping WebGPU support as a preview in Docker Desktop, but that it will be available in Docker Engine and other platforms in the future.
“What we’re doing is providing container images with multiple backends built in,” Cormack told The New Stack.
Firstly, this should make things easier for developers who have few options when it comes to the GPU on their development machine, but ultimately it may make things go smoother when their application goes into production.
“Our next step is to show you what that path looks like,” he said, “so you have a pre-built model that you can run locally or you can run in production.”
Increased access and openness
Cormack said the company is also looking to work with Llama Models: “We’re excited to work with this broad community to provide access to GPUs to anyone developing or working at the edge in a variety of environments, improving affordability and ease of use for the ecosystem.”
The conference highlighted the potential of open source AI in increasing access to the technology and bringing transparency to a market that can be dominated by large tech companies, but also raised concerns about “open washing.”
Ibrahim Haddad, executive director of the Linux Foundation AI and Data Foundation, announced tools to enable AI model creators (and end users) to evaluate how open their models actually are.
The isitopen.AI site is based on the Model Openness Framework developed by the Foundation, which divides models into 17 different components spanning code, tools, data and documentation.
The models are divided into three classes.
Class I – Open Science covers the full range of components with all artifacts made public, including training datasets, making the model fully reproducible. Class II – Open Tools covers the full suite of code and key datasets. At the bottom, Class III – Open Models means that the core of the model, including its architecture, parameters and basic documentation, is made publicly available under an open license.
Haddad said the complexity of AI compared to traditional code makes it difficult to maintain a binary approach to openness, but that incremental transparency is better than no transparency at all.
In a more practical sense, this tool and framework will enable developers and companies considering using the models to understand exactly how a particular model will mesh with their own frameworks and internal processes. It will also give model developers a clearer picture of their own status and ensure they do not use inappropriate licenses. It is noteworthy that none of the models analyzed so far have achieved Class III status or higher.
YOUTUBE.COM/THENEWSTACK
Technology is evolving quickly, so don’t miss an episode. Subscribe to our YouTube channel to stream all our podcasts, interviews, demos, and more.
subscribe
The group was created in Sketch.
Joe Fahy has covered the technology industry for 30 years, edited publications in London and San Francisco, and is a contributing analyst to GigaOm.
